The following results are from our simulation experiments on this platform, not from the published NEMO papers. They are predictions of the Assembly Calculus that we have validated computationally but not yet published.
A central finding in neurolinguistics is that the brain processes language in stages: initial structure building (~100–300ms, reflected in the ELAN), lexical-semantic integration (~300–500ms, the N400), and syntactic reanalysis (~500–1000ms, the P600). Friederici (2002) proposed a three-phase model based on ERP evidence. We find that a single Assembly Calculus architecture — with no separate modules for semantics or syntax — produces three independent metrics that map onto these phases, each selectively sensitive to its corresponding violation type. This is a triple dissociation: three measures from one system, each affected by a different manipulation.
The three metrics
All three metrics are measured during the same simulation run. They come from different brain areas and different moments in the process:
| Friederici phase | ERP analogue | Latency | Assembly Calculus metric | Where measured |
|---|---|---|---|---|
| Phase 1: Initial structure building | ELAN | ~100–300ms | Core-area instability (Jaccard distance between successive winner sets) | NOUN_CORE during word settling |
| Phase 2: Lexical-semantic integration | N400 | ~300–500ms | Global pre-k-WTA energy (sum of all synaptic inputs before top-k selection) | NOUN_CORE after context integration |
| Phase 3: Syntactic reanalysis | P600 | ~500–1000ms | Structural-area instability (Jaccard distance in structural areas across settling rounds) | ROLE_AGENT, ROLE_PATIENT, VP during structure binding |
Instability is the Jaccard distance between the set of winning neurons on successive settling rounds. When an assembly converges quickly (the same neurons win each round), instability is low. When the assembly — different neurons win on different rounds — instability is high. This oscillation occurs when the network has not formed a stable attractor for the current input, which happens when the input violates the patterns learned during training.
The dissociation pattern
The experiment tests three conditions: grammatical sentences, semantic violations (a noun where a different noun is expected, based on context), and category violations (a noun where a verb is expected). The prediction: each violation type should elevate its corresponding metric without affecting the others.
| Metric | Grammatical | Semantic violation | Category violation | Selectivity |
|---|---|---|---|---|
| Core-area instability | 1.33 (low) | 3.75 (high) | 2.03 (medium) | d = 32.7, p < 0.0001 |
| N400 energy (NOUN_CORE) | 23,537 (low) | 24,872 (high) | N/A (measured in VERB_CORE) | d = 7.9, p = 0.0001 |
| Structural instability | 1.14 (low) | 1.15 (null) | 1.59 (high) | d = 5.7, p = 0.0002 |
The critical comparison: semantic violations elevate N400 energy (d = 7.9) but leave structural instability unchanged (d = 0.11, p = 0.814). Category violations elevate structural instability (d = 5.7) but are measured in a different core area for N400. Core-area instability is highest for semantic violations, reflecting the difficulty of settling the lexical assembly for an unexpected word. Each metric responds to its designated violation type.
Per-area breakdown
The structural instability metric aggregates across structural areas, but the per-area breakdown reveals an important pattern: consolidated areas (those with Hebbian-strengthened pathways from training) show massive differentiation, while unconsolidated areas (those that have only random baseline connectivity) do not.
| Area | Consolidated? | Grammatical | Category violation | Cohen's d |
|---|---|---|---|---|
| ROLE_AGENT | Yes | 0.13 | 1.54 | 39.1 |
| ROLE_PATIENT | Yes | 0.08 | 1.57 | 26.2 |
| VP | Yes | 0.05 | 0.15 | 4.4 |
| SUBJ | No (random only) | 2.42 | 2.29 | −2.5 |
ROLE_AGENT and ROLE_PATIENT show the clearest dissociation: near-zero instability for grammatical inputs (the trained pathways guide the assembly to a stable attractor immediately) and high instability for category violations (the unexpected word class does not activate the trained pathways, so the assembly oscillates). SUBJ, which has only random baseline connectivity, shows high instability in both conditions — it has no trained attractor to converge to.
This per-area pattern depends on a specific implementation detail: bootstrap connectivity. Before training, the network materializes random baseline weights for all inter-area connections. Training then consolidates specific pathways via Hebbian plasticity without resetting the baseline. The result is that trained areas have strong attractors superimposed on a random substrate, while untrained areas have only the random substrate. The P600 (structural instability) is specifically a property of the consolidated areas — it measures the failure to activate a learned structural representation.
Morphosyntactic agreement
Category violations (wrong word class) are one type of syntactic error. Morphosyntactic agreement violations — such as “the cats runs” (number disagreement) — are another. In the model, these produce a different metric: elevated VP assembly distance (Cohen's d ≈ 1.5, p < 0.0001). Assembly distance measures how much the VP assembly's composition changes (which specific neurons are in the winning set), as opposed to instability (which measures how much the assembly oscillates between rounds).
This suggests the P600 has multiple computational substrates: category violations drive (the assembly cannot settle), while agreement violations drive representational shift (the assembly settles to a different configuration than expected). Both are detectable, but through different metrics. This distinction is not visible in human EEG — the P600 is a single scalp-recorded component — but it generates a prediction: the topographic distribution or source localization of the P600 should differ between category and agreement violations.
Why this matters
A triple dissociation from a single architecture is a strong constraint on any theory of language processing. Many models can produce one of these effects (N400-like semantic sensitivity is common in connectionist models). Producing all three — with selective sensitivity, from the same dynamics, without dedicated modules — is much harder. The Assembly Calculus achieves it because the three metrics are natural properties of different aspects of the same recurrent Hebbian settling process: early settling difficulty (core instability), aggregate input cost (pre-k-WTA energy), and late structural convergence failure (structural instability).
The connection to Friederici's (2002) three-phase model is specific: Phase 1 (initial structure building, ~100–300ms, the ELAN) maps to core-area instability, Phase 2 (lexical-semantic integration, ~300–500ms, the N400) maps to pre-k-WTA energy, and Phase 3 (syntactic reanalysis, ~500–1000ms, the P600) maps to structural-area instability. The model does not implement these phases as sequential stages — they emerge from the dynamics of settling in different brain areas with different connectivity histories.